Container Images Considered Harmful
summary
Argues that tar-based OCI/Docker images cause wasted transfer and storage by forcing file-level or whole-layer copies when metadata or a single byte changes and by making builds noncanonical. Proposes chunk-level content defined chunking plus a content-addressable Merkle tree so identical chunks deduplicate across images, canonical JSON metadata for reproducible hashes, and using kernel features like copy_file_range / reflinks for low-cost on-disk sharing. Shows why these choices speed pulls, shrink storage, and make image contents auditable.
AI-written summary · about